PSI - Issue 62

Lorenzo Principi et al. / Procedia Structural Integrity 62 (2024) 89–96 Principi L. / Structural Integrity Procedia 00 (2019) 000 – 000

94

6

Fig. 1. Details of the k-fold cross-validation

Table 4. Hyper-parameters and Tested Values Hyper-parameter

Tested Values

Number of hidden layers

1, 2, 3, 4, 5

Number of neurons for hidden layers

12, 26

Optimization algorithm

Limited-memory BFGS, Stochastic Gradient Descent, ADAM Logistic, Rectified Linear Unit, Hyperbolic Tangent

Activation function

Alpha value

0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0

Max iterations number

10, 20, 30, 40, 50, 100, 150, 200, 250, 300, 350, 400, 450, 500

* Batch size

4, 8, 16, 32

* Initial learning rate Number of hidden layers

0.1, 0.01, 0.001, 0.0001

1, 2, 3, 4, 5

Number of neurons for hidden layers * SGD and ADAM solvers only

12, 26

Table 5 . Best model’s architecture and values of hyperparameters Hyperparameter Value Number of hidden layers 3 Number of neurons 26 per layer Optimization algorithm LBFGS Activation function Hyperbolic Tangent Alpha value 0.20 *2 - 0.30 *1 Max iterations number 50 *2 – 30 *1

*1 For LoD; *2 For Structural and Foundational Attention Class

5. Results This section discusses the results obtained in predicting both the Level of Defectiveness and the Structural and Foundational Attention Class on the case study. Fig. 2a presents the confusion matrix, and Table 6 provides the classification report for Level of Defectiveness. The average F1-Score obtained is 0.51 (Accuracy 0.58). The ANN can adequately predict the Medium-Low to Medium-High Classes, achieving an F1-Score between 0.55 and 0.71. However, there is room for improvement in predicting the other classes, especially for the Low and High Classes, where F1-Scores are 0.22 and 0.38, respectively. Fig. 2b presents the confusion matrix, and Table 6 provides the classification report for the Structural and Foundational Attention Class. The overall F1-Score achieved is 0.77 (Accuracy 0.77). On average, the model performs well in predicting the Low, Medium-High and High Classes, with F1-Scores of 1.00, 0.83 and 0.86, respectively. Nevertheless, refinements are needed to improve the prediction of both